Post on 06-Aug-2020
AN INFINITE DIMENSIONAL CENTRAL LIMIT THEOREMFOR CORRELATED MARTINGALES
ILIE GRIGORESCU
Abstract. The paper derives a functional central limit theorem for the em-
pirical distributions of a system of strongly correlated continuous martingales
at the level of the full trajectory space. We provide a general class of function-
als for which the weak convergence to a centered Gaussian random field takes
place. An explicit formula for the covariance is established and a characteri-
zation of the limit is given in terms of an inductive system of SPDEs. We also
show a density theorem for a Sobolev-type class of functionals on the space of
continuous functions.
1. Introduction
The classical law of large numbers for empirical measures states that, given a se-
quence of independent random variables Z1, Z2, . . . with values in a Polish space
E and common distribution α(dx) ∈ M(E), the random measures N−1∑N
i=1 δZi
converge weakly in probability to α(dx) as N → ∞. Furthermore, the fluctuation
random field ξN = N− 12
∑Ni=1(δZi − α(dx)) converges to a centered Gaussian ξ in
the sense that, for any test function g ∈ Cb(E), the space of bounded continuous
functions on E,
(1.1)(g,
1√N
N∑
i=1
(δZi − α(dx))
converges in distribution to a normal random variable with mean zero and variance
σ2ξ (g, g) = Covα(g, g).
We are interested in deriving a central limit theorem when E is the path space
Ω = C([0, T ],R) up to time T > 0 and the random variables Z1, Z2, . . . are replaced
Date: May 11, 2003.
1991 Mathematics Subject Classification. Primary: 60F17; Secondary: 60G15, 60K35.
Key words and phrases. Gaussian random field, fluctuations from hydrodynamic limit, central
limit theorem.
Research supported in part by McMaster University and The Fields Institute, Ontario, Canada,
the University of Utah and the University of Miami.
1
2 ILIE GRIGORESCU
with correlated processes. More precisely, the random field (1.1) can be calculated
for the random variables Zi = zNi (·), designating the trajectory of a Markovian
system of coupled particles at times t ∈ [0, T ], which can be viewed as path valued
random variables. A first step is to extend the notion of fluctuation random field
to functionals of the path ω(·) ∈ C([0, T ],R) up to the time horizon T > 0 of the
particles, that is, at a minimum, to functionals depending of finitely many time
marginals G(ω(·)) = g(ω(t0), ω(t1), . . . , ω(tm)) for some positive integer m ≥ 1 and
g smooth.
The central limit theorem for empirical measures is well known for independent
random variables and has been studied as a distribution valued continuous process
for the case of Brownian motions by Ito in [8]. In [7], Holley and Stroock introduce
the theory of generalized Ornstein-Uhlenbeck processes and prove the central limit
theorems for various interacting particle systems. Other limits concerning fluctu-
ation random fields from the hydrodynamic profile of interacting particle systems
can be found in [4] for zero range processes in equilibrium, a non-equilibrium re-
sult for symmetric simple exclusion appears in [14] and [2] solves the problem for
Ginzburg-Landau lattice models. In all these results the limiting random field is a
time-indexed continuous Markovian process with values in the space of tempered
distributions, that is, can be obtained for the special case m = 0 and g ∈ C∞(R)
of Schwartz class.
The result which appears to be the closest in spirit to the present work is [16].
The limit is an infinite dimensional random field, but the coefficients individually
satisfy a one-dimensional central limit theorem, a feature due to the mean field
character of the model.
Because of the strong correlations (2.8) we cannot keep the bounded continuous
functions on Ω as index set. A natural answer is to use the class of functions
with bounded smooth derivatives C1b (Ω) given in Definition 1. A price paid for the
generalization is that we adopt an example of correlated continuous martingales
described by (2.8) in order to have access to concrete calculations. However, this
example provides an additional benefit with the derivation of an explicit formula for
the covariance function (2.12). The paper has an important constructive component
since Sections 2, 3 and the imbedding Theorem 5 are laying the ground for an infinite
dimensional result in a general setting. Here they are used in establishing the main
result, Theorem 3.
AN INFINITE DIMENSIONAL CENTRAL LIMIT THEOREM 3
The immediate motivation of the present work is the scaling limit for the Brow-
nian motions with local interaction on the unit circle S1 from [5]. After calculat-
ing explicitly the asymptotic law of a single particle in the random environment
provided by the rest of the interacting system (the tagged particle problem) and
establishing that finite subfamilies of particles become independent in the scaling
limit (propagation of chaos) one has an immediate weak law of large numbers for
the empirical measures associated to the process at the level of the full trajectory
space C([0, T ], S1). The process has a product uniform invariant measure and the
actual law of the tagged particle process in that case is the Wiener measure on the
unit circle with uniform initial distribution. A natural question is whether one can
determine the limit for the fluctuation field from the mean, at least in equilibrium.
There is another reason why we need a result at the level of the full path space.
Due to the symmetry of interacting diffusions in [5], the hydrodynamic limit and
the fluctuation random field from the weak solution to the heat equation describing
the trajectory of the empirical measures indexed by time are the same as in the case
of independent Brownian motions. The quantities present are, from our viewpoint,
just one-dimensional marginals of the objects we are interested in. The interaction
surfaces only at the level of the path space, or when correlations between config-
urations observed at consecutive times are taken into account (the history of the
process). In a different formulation, the interaction becomes apparent when we
consider a multi-color version of the process, as coloring is a weak form of tagging
(see, in that sense, the comments in [6] and the approach used in [12] and [13] for
symmetric simple exclusion).
The proof of this particle model offers a hint into the nature of the scaling needed
for the fluctuation field. Through a path transformation, the system of particles
can be converted into a family of martingales on the Skorohod space, adapted to
the filtration of the original process. Since the correlations are of the order of the
inverse of the number of particles, they survive in the limit. This provides the
correct scaling limit (which is also the natural one from the classic CLT).
Unfortunately, even though the path transformation has a smooth asymptotic
value which is invertible pathwise, it is discontinuous before the limit and the error
from the continuum limit is once again of the order of the square root of the number
of particles, a finite but still too large a perturbation in order to establish a CLT
like Theorem 3. For our present purposes, one can summarize the example of the
induced martingales with the following construction.
4 ILIE GRIGORESCU
Let λ and ρ be two positive constants. For every N ∈ Z+ we consider a set of
N2 independent Brownian motions on a probability space (W,F , P ), adapted to
a filtration Ftt≥0, as follows. We shall have, for 1 ≤ i ≤ N a collection βi(·),plus another family of N(N − 1) independent Brownian motions wij(·), with 1 ≤i 6= j ≤ N such that βi(0)1≤i≤N are i.i.d. with common probability distribution
µ(dx) and wij(0) = 0 P -almost surely. Let
(1.2) zNi (t) = βi(t) +
(ρ
λ + ρ
)1N
∑
j 6=i
(βj(t)− βi(t))+
+( √
λρ
λ + ρ
)1√2N
∑
k 6=i
(wik(t)− wki(t)) .
The quadratic variation is
〈zNi , zN
i 〉(t) =λ
λ + ρt + O(
1N
)t
and the cross variation processes for i 6= j are
〈zNi , zN
j 〉(t) =1N
(ρ
λ + ρ
)t + O(
1N2
)t .
It is clear that the construction satisfies assumptions (2.13) and (2.14). The martin-
gales generated in the original problem emulate the interacting diffusions very well
and in equilibrium they have identical limit of the empirical measures (in the sense
of Theorem 1). In the context of [5], the parameter ρ represents the average density
of the particles on the unit circle and λ controls the intensity of the interaction.
The present paper determines the fluctuation limit in the case of correlated
Brownian motions like in the example from above. The discussion following the
main results Theorems 2 and 3 and especially Remark 2 after Theorem 3 are helpful
in completing the present discussion.
Naturally under weak conditions (Assumption 1), a finite-dimensional CLT will
hold, that is, for smooth cylinder test functions. It is remarkable that the covariance
can be given in a closed explicit formula (2.12). This fact can be generalized if the
limiting one-particle process Q has a time-only dependent generator, but cannot be
done along the same lines in the presence of path or spatial coordinate dependence.
The passage from cylinder functions to a convenient larger space needs much more
stringent conditions (Assumption 2), as it can be seen from Section 5.
Even though the correlated martingales are far from the complexities occurring
in interacting particle system, one hopes that, at least in equilibrium, formula (2.12)
will be the same. The function space C1b (Ω) introduced in Definition 1 is a natural
AN INFINITE DIMENSIONAL CENTRAL LIMIT THEOREM 5
candidate for test functions if only we look at the one-dimensional case, for instance.
Also, the inductive construction presented in Sections 3 and 4 can be adapted to
essentially any Markovian model.
We give an outline of the way the paper is organized. Section 2 introduces the
spaces of cylinder functions, in particular the special cylinder functions of exponen-
tial or Schwartz type E(Ω), respectively S(Ω), in Definition 3. The Banach space
H(Ω, Q) presented in Definition 6 allows us to link the space of cylinder functions
and the space of test functions C1b (Ω) (Definition 1) over which Theorem 3 (the
infinite dimensional case) is established through an imbedding result - Theorem 5,
Section 5. In addition, a few examples of relevant test functions are provided in
the remarks after Definition 6 and further down in Proposition 2.
Section 3 lays out an inductive characterization of the random fields through
Theorem 4, paired with an inductive characterization of the covariance function in
Proposition 3.
Propositions 5 and 6 in Section 4 prove Theorem 2, the central limit theorem
for cylinder functions in (S ∪ E)(Ω). In particular, Proposition 5 and 6 show that
the one-dimensional marginal of the limiting Gaussian random field ξ is a time-
continuous distribution - valued Markov process solving an Ornstein-Uhlenbeck
SPDE, which is consistent with [7] and [8]. Because S(Ω) is a linear space we
obtain that, for any linear combination∑
l clGl of functions Gl ∈ S(Ω), the random
variables∑
l cl(Gl, ξN ) converge to a Gaussian (
∑l clGl, ξ). The actual covariance
of the limit is obtained after matching the inductive characterization with the actual
solution, which is done in Proposition 4. This identification offers an example of
a nontrivial solution of the inductive SPDE associated to the half Laplacian and a
specific bilinear form q(·, ·) in the sense of (3.4) in Definition 14.
Section 5 is based on Theorem 5 proved in the Appendix and the asymptotic
uniform bound (5.7) from Proposition 9. The latter needs (2.13) - (2.14) from
Assumption 2 in order to complete a martingale representation (2.15) through the
series of Lemmas 1 - 2.
Finally, the Appendix proves non-probabilistic results generally valid in function
spaces as well as Theorem 1 which is a trivial hydrodynamic limit in this context.
6 ILIE GRIGORESCU
2. Definition and results
Let Ω = C([0, T ], X) be the space of continuous paths up to time T > 0 on X
which will be either the real line R or the unit circle S1. The uniform norm on X
will be denoted by ‖ · ‖.
Definition 1. Let C be the set of complex numbers and let G ∈ C(Ω) = C(Ω,C) be
the space of complex valued continuous functions and ω ∈ Ω be fixed. Assume there
exists a continuous linear mapping η → ∇ηG(ω), for all η ∈ Ω, and a function
c(G,h) depending only on G and h such that, for any η ∈ Ω and h ∈ R
(2.1) |G(ω + h η)−G(ω)− h∇ηG(ω)| ≤ c(G,h)‖η‖2
with limh→0(c(G, h)/h) = 0. The Frechet derivative will have the strong norm
(2.2) ‖∇·G(ω)‖ = supη 6=0
|∇ηG(ω)|‖η‖ .
We shall say that G ∈ C1b (Ω) if G(ω), ∇·G(ω) are uniformly bounded on Ω. The
space C1b (Ω) is a normed linear space with the norm
(2.3) ‖G‖C1b
= supω∈Ω
(|G(ω)|+ ‖∇·G(ω)‖) .
The mappings ∇·G(ω) are signed measures on [0, T ] depending on ω and the norm
defined by (2.2) is the total variation norm.
Definition 2. We shall denote by E(R) the set all exponential functions of the form
x → eiαx, with α ∈ R, by C(R) the set of bounded infinitely differentiable functions
on R with bounded derivatives and by S(R) the Schwartz class of functions rapidly
decreasing at infinity.
Definition 3. The functional G ∈ C1b (Ω) on the path space Ω will be said to be
C - class cylinder function on Ω (respectively of S - class or E - class) if there
exist a positive integer m, an increasing sequence of times 0 ≤ t1 < t2 < . . . < tm
and a family of functions gl(x) ∈ C(R) (respectively of S - class or E - class),
1 ≤ l ≤ m such that G(ω) = Πml=1gl(ω(tl)). The space of such functions will be
denoted by Ccyl(Ω) and the linear span of such functions will be denoted by C(Ω).
In the same way, the space (E ∪ S)(Ω) is the linear span of cylinder functions with
factors belonging either to E(Ω) or S(Ω).
For G ∈ Ccyl(Ω) we define a(ω(·)) = Πm−1l=1 gl(ω(tl)) and t′ = tm−1 ≥ 0. We denote
Ms = σ[ω(u) : 0 ≤ u ≤ s] the σ - algebra generated by the continuous paths up
AN INFINITE DIMENSIONAL CENTRAL LIMIT THEOREM 7
to time s ∈ [0, T ]. Then a(ω(·)) is a Mt′ - measurable functional on Ω. We shall
look at the test function a(ω(·))g(ω(t)) for t ≥ t′ which we shall call the associate
marginal process of G starting at t = t′. We denote by
(2.4)G(ω, t) = a(ω(·))g(ω(t)) , t ≥ t′ = tm−1 ,
∂G(ω, t) = a(ω(·))g′(ω(t)) and ∂2G(ω, t) = a(ω(·))g′′(ω(t))
the (inductive) cylindrical decomposition of G and its derivatives. In the same time,
any G ∈ C(Ω) can be written as the sum of functions from Ccyl(Ω) with the same
m > 0 by considering the union of all times t from all the terms in G and formally
factoring in some constant functions (provided that tm is indeed the largest time
present in G). This allows us to extend the definitions (2.4) to any G ∈ C(Ω).
We consider a probability space (W,F , P ), where F is a filtration Ft0≤t≤T on
W and take σ2 > 0.
Definition 4. We shall denote by Q the family of laws of the diffusion processes
Qν on Ω with respect to (W,F , P ), adapted to an extension of Ftt≥0, with genera-
tor σ2
2d2
dx2 and initial distribution ν(dx), where ν(dx) is a probability measure on X.
In the following, Q0 will designate the Brownian motion with diffusion coefficient
σ2 starting at zero and Q = Qµ (without superscript) for simplicity.
Definition 5. For G ∈ C1b (Ω) we define the linear functional on Ω
(2.5) η −→ 〈∇ηG〉Q =∫
Ω
∇ηG(ω)dQ(ω) .
Remark: This definition is consistent pointwise since the gradient of G is bounded
by ‖η‖. The law Q of the diffusion is the Wiener measure on the path space Ω
and ω(·, w) is a random variable measurable with respect to (W,F), distributed
according to Q.
Definition 6. We shall denote by H(Ω, Q) the Banach space obtained by completion
of the space C1b (Ω) under the norm ‖ · ‖H defined as
(2.6) ‖G‖2H =∫
Ω
|G(ω)|2dQ(ω) +∫
Ω
‖∇·G(ω)‖2dQ(ω) .
Remark 1: H(Ω, Q) is not a Hilbert space. However, for finite-dimensional
marginals, that is for cylinder functions g with m ∈ Z+, we obtain H1(Rm+1).
Remark 2: Proposition 2 provides a general class of examples of functionals in
H(Ω, Q) (cylinder functions). Also, an important case of test function G which
8 ILIE GRIGORESCU
belongs to H(Ω, Q) but not to S(Ω) is G(ω(·)) = ω(t), the projection at a given
time t. See also the remark after Corollary 1.
Remark 3: If b(·, ·) ∈ C0,2([0, T ], X), the functional G(ω) =∫ t
0b(s, ω(s))ds be-
longs to C1b (Ω).
Proposition 1. The linear functional 〈∇ηG〉Q is square integrable with respect to
Q0 and
(2.7)∫
Ω
|〈∇ηG〉Q|2dQ0(η) ≤ (2σ2T )‖G‖2H .
Proof:
|〈∇ηG〉Q|2 =∣∣∣∣∫
Ω
(∇ηG(ω)‖η‖
)‖η‖dQ(ω)
∣∣∣∣2
≤∣∣∣∣∫
Ω
‖∇·G(ω)‖‖η‖dQ(ω)∣∣∣∣2
≤(∫
Ω
‖∇·G(ω)‖2dQ(ω))‖η‖2 .
We can take the expected value with respect to Q0(η) and obtain (2.7) using Doob’s
inequality (see in [15]). ¤We would like to know how large H(Ω, Q) is. The set S(Ω) of Schwartz class
cylinder functions is a subset of the set of smooth cylinder functions C(Ω) which is
included in H(Ω, Q). Theorem 5 and Proposition 7 will show that S(Ω) is dense in
H(Ω, Q).
Proposition 2. Let m ∈ Z+ and g(x0, x1, . . . , xm) be a function in the space
H1(Rm+1), that is with g and its derivative in the sense of distributions square
integrable with respect to the Lebesgue measure. If µ(dx), the initial distribution of
Q, is absolutely continuous with respect to the Lebesgue measure and the density
ρ(x) defined as µ(dx) = ρ(x)dx is bounded, then G(ω) = g(ω(t0), . . . , ω(tm)) ∈H(Ω, Q). Also, if G does not depend on the initial time t = 0 the statement is valid
for arbitrary µ(dx) as long as G depends on a finite number of times.
Proof of Proposition 2: Any function g ∈ H1(Rm+1) can be approximated by
gS ∈ C(Rm+1) in the H1 norm. The statement is proven if we can show it for a
Schwartz class function. The preceding lemma has shown that, for any two ω, η ∈ Ω
we have
∇ηG(ω) =m∑
i=0
(∂xig(ω(t0), . . . , ω(tm)))η(ti) .
We see that
|∇ηG(ω)| ≤ (m∑
i=0
(∂xig(ω(t0), . . . , ω(tm)))2)12 (
m∑
i=0
(η(ti))2)12 ≤
AN INFINITE DIMENSIONAL CENTRAL LIMIT THEOREM 9
≤ (m + 1)12 ‖∇xg(ω(t0), . . . , ω(tm))‖Rm+1‖η‖ .
The conditions from the proposition make the joint probability density function
of the random variables (ω(t0), . . . , ω(tm)) be a bounded function on Rm+1. The
formula for the joint density is[µ(y0)Πm
i=1p(ti − ti−1, yi − yi−1)]dy0dy1 . . . dym
with p(t, y) the kernel of the heat equation ut = (σ2/2)uxx. Since the heat equation
semigroup produces smooth bounded functions for any t > 0 irrespective of the
initial distribution µ(dx) the inequality p(ti − ti−1, yi − yi−1) ≤ Const σ−1(ti −ti−1)−
12 for i ≥ 2 finishes the proof. ¤
We need to introduce the general setting for Gaussian processes. Let X be a
Banach space with norm ‖ · ‖X . For any complex number z we denote by z its
complex conjugate.
Definition 7. A continuous bilinear form on X is a mapping q : X ×X −→ C such
that, for any F, G ∈ X , q(·, G) and q(F, ·) are linear and there exists a constant
c(q), independent from F and G, such that |q(F,G)| ≤ c(q)‖F‖X ‖G‖X . We shall
write q ∈ B(X ). The bilinear form is called symmetric if q(F, G) = q(G,F ) ,
nonnegative if q(G,G) ≥ 0 and positive if q(G,G) > 0 if G 6= 0.
A random process indexed by A is, by definition, a collection of real-valued
random variables ξ = ξaa∈A on a probability space (W,F , P ). The sub-algebra
Σ generated by
Ca1,a2,...,an,B = w ∈ W : (ξa1(w), . . . , ξan(w)) ∈ B ∈ F ,
where n ∈ Z+, a1, a2, . . . , an ∈ A and B ∈ B(Rn) are arbitrary, allows us to define
a probability measure on RA with the σ - field generated by the finite dimensional
projections. For any finite set A0 = a1, a2, . . . , an ⊆ A we define the probability
measure on RF , called the finite dimensional distribution of ξ on A0 by
PA0(B) = P ((ξa1(w), . . . , ξan(w)) ∈ B) .
The consistency condition of the finite-dimensional distributions is respected. For
any two finite subsets of A such that A′0 ⊆ A0 we define πA′0 the projection of RA0
onto RA′0 and then
PA0 π−1A′0 = PA′0 .
10 ILIE GRIGORESCU
Kolmogorov’s extension theorem shows that under these circumstances there exists
a probability measure on RA and the σ - field generated by the finite dimensional
cylinder functions denoted by P such that, if F is a finite subset of A and πA0 is
the projection of RA onto RA0 , then
P π−1A0
= PA0 .
Definition 8. Let ξN = ξNa a∈A defined for all N > 0 and a separate ξ = ξaa∈A
be random processes indexed by A. We shall say that ξN converges to ξ as N →∞ if
the finite dimensional distributions of ξN converge weakly to the finite dimensional
distributions of ξ.
Definition 9. The random process ξ indexed by A is called Gaussian (centered
Gaussian) if all linear combinations of ξa with a ∈ A are Gaussian (centered Gauss-
ian).
Definition 10. Let X be a Banach space. The random process ξ indexed by F ∈ Xwith the property that the mapping F → ξF belongs to X ′, the space of linear
functionals on X , is called a random field on X and we shall write ξF = (F, ξ). In
case the random process defined this way is Gaussian (centered Gaussian) we shall
say that ξ is a Gaussian (centered Gaussian) random field.
For X equal to the space H(Ω, Q) and q(·, ·) a nonnegative continuous symmetric
bilinear form on H(Ω, Q), there exists a centered Gaussian random field ξ = ξ(w)
on X with covariance q(F − 〈F 〉Q, G − 〈G〉Q) for any two F, G ∈ H(Ω, Q). Here
and in the following 〈G〉Q =∫Ω
GdQ(ω). In this case we shall denote the variance
by σ2ξ (·, ·).
We are now ready to formulate the assumptions needed for our results. For
every positive integer N , we consider a family of N continuous square-integrable
martingales with respect to P and Ft0≤t≤T , taking values in X, denoted by
zNi (·, w)1≤i≤N . The cross variation processes 〈zN
i , zNj 〉(t,w)t≥0, 1 ≤ i, j ≤ N
can be written as
(2.8) 〈zNi , zN
i 〉(t,w) =∫ t
0
σ2N,i(s, w)ds and 〈zN
i , zNj 〉(t,w) =
∫ t
0
kijN (s,w)ds ,
where σ2N,i(·,w) and |kij
N (·, w)| are time-integrable a.s. with respect to P .
AN INFINITE DIMENSIONAL CENTRAL LIMIT THEOREM 11
Equivalently, by the martingale representation theorem, possibly by extending
the σ-field F , one can define the N -dimensional Brownian motion z(·, w) by
dzi(t,w) =L∑
l=1
ril(t, w)dwl(t,w) , 1 ≤ i ≤ N , 1 ≤ l ≤ L
where L ∈ Z+ and w(·, w) = wl(·, w)1≤l≤L is an L-dimensional Brownian mo-
tion adapted to F and R(t, w)R∗(t, w) = (∑
l ril(t, w)rlj(t,w))ij is the correlation
matrix of z(·,w) with elements given by the integrands from (2.8).
In order to derive a finite-dimensional central limit theorem for the empirical
measures associated to the family of martingales zNi (·)1≤i≤N we shall only need
the following condition.
Assumption 1. (Finite dimensional CLT) There exist σ2 > 0 and γ > −1
such that
(2.9) limN→∞
E[ ∫ T
0
N∑
i=1
(σ2N,i(s, w)− σ2)2ds
]= 0
and
(2.10) limN→∞
E[ ∫ T
0
1N2
∑
1≤i6=j≤N
(NkijN (s, w)− σ2γ)2ds
]= 0 .
Theorem 1. Under Assumption 1, if there exists a probability measure µ(dx) on X
such that the initial empirical measures N−1∑N
i=1 δzNi (0) converge weakly to µ(dx)
in probability, then the empirical measure N−1∑N
i=1 δzi(·) converges weakly to Q in
probability.
The proof of Theorem 1 is in the Appendix.
Theorem 2. Under Assumption 1, where (2.9) and (2.10) are satisfied with γ > −1
and the initial values of the martingales zN1 (0), . . . , zN
N (0) are independent with
distribution µ(dx) on X, the random field
(2.11) ξN =1√N
N∑
i=1
(δzNi (·) −Q)
on Scyl(Ω) converges in the sense of Definition 8 to a centered Gaussian random
field ξ on Scyl(Ω) with covariance
(2.12)σ2
ξ (F, G) =∫Ω(F (ω)− 〈F 〉Q)(G(ω)− 〈G〉Q)dQ(ω)
+γ∫Ω〈∇ηF 〉Q〈∇ηG〉QdQ0(η)
for any F,G ∈ Scyl(Ω).
12 ILIE GRIGORESCU
A stronger set of condition is adopted for an infinite dimensional central limit
theorem.
Assumption 2. (Infinite dimensional CLT) There exist σ2 > 0 and γ > −1,
as well as constants C(σ,N), C(γ,N), for all N ≥ 1 independent from w ∈ W and
t ∈ [0, T ], with the property limN→∞ C(σ,N) = 0 and limN→∞ C(γ, N) = 0 such
that properties (2.13) and (2.14) are valid P -almost surely:
(2.13)√
N max1≤i≤N
[sup
0≤t≤T
∣∣σ2N,i(t,w)− σ2
∣∣]≤ C(σ,N)
and
(2.14) max1≤i,j≤N
[sup
0≤t≤T
∣∣∣NkijN (t, w)− σ2γ
∣∣∣]≤ C(γ, N) .
Remark 1: Lemma 1 ensures the existence of a system of martingales with these
properties, in other words, that the covariance matrix of the martingales stays
positive definite.
Remark 2: Assumption 2 implies Assumption 1.
Remark 3: In the following we shall omit the random element w when not neces-
sary.
Theorem 3. Under Assumption 2, where (2.13) and (2.14) are satisfied with γ >
−1 and the initial values of the martingales zN1 (0), . . . , zN
N (0) are independent with
distribution µ(dx) on X, the random field (2.11) on C1b (Ω) converges in the sense of
Definition 8 to a centered Gaussian random field ξ on C1b (Ω) with covariance (2.12)
for any F, G ∈ C1b (Ω). Furthermore, since the covariance (2.12) is continuous with
respect to the norm ‖·‖H, the limit ξ can be extended to a centered Gaussian random
field on H(Ω, Q).
Remark 1: The random field ξN is in fact a random measure on Ω for any N > 0.
As a consequence, ξN has values in (H(Ω, Q))′ almost surely. However, one cannot
carry out the limit uniformly over G ∈ H(Ω, Q) to prove that the limit ξ has values
in (H(Ω, Q))′ even in the independent case. Still, if we drop the requirement that
the measure on (H(Ω, Q))′ be countably additive (no longer a measure in the proper
sense) we can define a so-called cylindrical measure (see [1], Section 3.9).
Remark 2: Suppose we keep the correspondence σ2 = λ/(λ + ρ) and γ = ρ/λ, in
view of the example (1.2). Then we can see that, in the strong interacting case when
λ → 0, the first term of the covariance (2.12), corresponding to the classical central
AN INFINITE DIMENSIONAL CENTRAL LIMIT THEOREM 13
limit theorem (noncorrelated case) will vanish, while the second part, corresponding
to the correlation, will tend to one. This can be seen because, as the measure
QN converges to a degenerate measure, the factor EQ0[(〈∇ηG〉Q)2] ∼ EQ0
[‖η‖2],which is of the order of λ. Multiplied by γ = ρ/λ we obtain a limit of order
one. This is natural when the particles are, in fact, moving deterministically. The
remaining randomness originates from the classical CLT for the initial positions of
the particles. On the other hand, the case of very rare particles ρ → 0 eliminates
correlation. The weak interaction case λ →∞ pushes the diffusion coefficient to one
(independent Brownian motions) and the asymptotic correlation vanishes again.
Theorem 3 extends the finite dimensional result of Theorem 2 to the infinite
dimensional space C1b (Ω). Theorem 5 imbeds the space of cylinder functions into
H(Ω, Q), providing a density theorem S(Ω)H ⊇ C1b (Ω), where the subscript desig-
nates the norm (2.6). In order to make use of this imbedding, we need an uniform
bound (also a tightness estimate) with respect to N on the second moments of
the random field ξN . This is done through (5.7) from Proposition 9. In order to
understand better Assumption 2 we have to write
(2.15) zNi (t) = yN
i (t) +1√N
dNi (t) , 1 ≤ i ≤ N
as in Lemma 2, where yNi (·)1≤i≤N are independent Brownian motions and the
continuous martingales dNi (·)1≤i≤N have asymptotically bounded moments of
some order r > 2. The construction can be viewed as an orthonormalization pro-
cedure. Assumption 2 is sufficient for this construction as well as for a certain
simplicity of the result. Following the proof of Proposition 9 and Lemmas 1 and 2
one can try to relax the assumptions (2.13) - (2.14) to
(2.16) E
∫ T
0
(σ2
N,i(s,w)− σ2)2
ds +∑
j 6=i
∫ T
0
(kij
N (s, w)− σ2γ
N
)2
ds
∼ o(
1N
)
uniformly in 1 ≤ i ≤ N . However, this weaker set of conditions complicates the
proof of Proposition 8. We shall not pursue this direction in the present paper.
3. Inductive characterization of a random field
The following definitions formulate the main conditions needed for the characteriza-
tion of a centered Gaussian random field by induction on the maximum number of
times m appearing in the test functions from C(Ω). We recall that Q is a Brownian
motion as in Definition 4.
14 ILIE GRIGORESCU
Definition 11. Let q(·, ·) ∈ B(C1b (Ω)) be symmetric, nonnegative and continuous
in the H(Ω, Q) norm as in Definition 7. For any pair of functionals F, G ∈ C1b (Ω)
of the form
(3.1) F (ω(·)) = f(ω(0)) and G(ω(·)) = g(ω(0))
with f, g ∈ C1b (X), the bilinear form q0(f, g) = q(F, G) is well defined, symmetric,
nonnegative and continuous with respect to the induced norm ‖·‖H(X,µ) on H(X, µ).
This one-dimensional bilinear form will be called the one-dimensional marginal of
q(·, ·) at t = 0.
Definition 12. A bilinear form u(·, ·) ∈ B(C1b (Ω)) is said translation invariant if
u(F + c, G) = u(F, G) and u(F,G + c) = u(F,G) for any c ∈ C.
Definition 13. Let q(·, ·) ∈ B(C1b (Ω)) be a positive symmetric bilinear form contin-
uous with respect to the H(Ω, Q) norm. We shall say that u(·, ·) solves inductively
the differential equation for the operator (σ2/2)∆ and the bilinear form q(·, ·) with
initial marginal at time t = 0 denoted by q0(·, ·) if
(3.2)
u(F (ω, t), G(ω, t))− u(F (ω, t′), G(ω, t′)) =
σ2
2
∫ t
t′
(u(∂2F (ω, s), G(ω, s))+
u(F (ω, s), ∂2G(ω, s)) + 2q(∂F (ω, s), ∂G(ω, s))
)ds
for all F, G ∈ E(Ω) and u0(F, G) = q0(F, G) as in Definition 11.
Proposition 3. Let q(·, ·) ∈ B(C1b (Ω)) be a positive continuous symmetric bilinear
form and q0(·, ·) is its one-dimensional marginal. If there exists a translation in-
variant bilinear form u(·, ·) on C1b (Ω) starting at q0(f −〈f〉µ, g−〈g〉µ) in the sense
of Definition 11 satisfying the inductive equation (3.2), then u(·, ·) is unique.
Remark 1: The inductive partial differential equation is valid for all smooth cylin-
der functions. What Proposition 3 and later Theorem 4 imply is that it is enough
to verify it for the E - class cylinder functions. Given that q(·, ·) is continuous in
the ‖ · ‖H norm, we can go from E(Ω) to S(Ω) and finally to H(Ω, Q).
Remark 2: Proposition 3 is valid if we assume equation (3.2) is verified for pairs
(F, G) with F = G.
Proof of Proposition 3: The difference v(·, ·) of two solutions of (3.2) is a translation
invariant bilinear form starting at zero and is a solution of the same equation (3.2)
AN INFINITE DIMENSIONAL CENTRAL LIMIT THEOREM 15
where the term in q(·, ·) is cancelled, namely a solution to
(3.3) v(F (ω, t), G(ω, t))− v(F (ω, t′), G(ω, t′)) =
σ2
2
∫ t
t′
(v(∂2F (ω, s), G(ω, s)) + v(F (ω, s), ∂2G(ω, s))
)ds .
We want to show that such a solution is zero. To prove this fact, we concentrate
on purely cylindrical functions and proceed by induction on the number of fac-
tors present. For all G ∈ S(Ω) one can write G(ω(·)) =∑
l Gl(ω(·)), where we
made the assumption that the summation runs through a finite set, with Gl ∈Scyl(Ω). Any function Gl can be written, according to (2.4), in the form Gl(ω, t) =
aG,l(ω(·))gG,l(ω(t)) where t ≥ t′ is the largest time variable present in G and t′ is
the next time smaller than t. This implies that
G(ω, t) =∫
R
∑
l
[(aG,l(ω) exp (iαω(t)))gG,l(α)]dα ,
for α ∈ R and the analogous formulas are valid for F ∈ S(Ω). Then, we replace
F (ω, t) by∫R aF (ω)eiαω(t)gF (α)dα and G(ω, t) by
∫R aG(α)eiαω(t)gG(α)dα where
(α, α) ∈ R2.
We first prove the uniqueness in the case of Fα(ω, t) = aF (ω)eiαω(t) and Gα(ω, t) =
aG(ω)eiαω(t). We look at (3.3) as an ordinary differential equation with unknown
q(Fα(ω, t), Gα(ω, t)). We can calculate explicitly the solution which is unique and
equal to zero as long as the induction hypothesis on the values at time t = t′ is
satisfied. The solutions will be bounded and hence we can write the full formula in
terms of Fα(ω, t) = aF (ω)eiαω(t) and Gα(ω, t) = aG(ω)eiαω(t) as
v(F (ω, t), G(ω, t)) =∫
R
∫
Rv(Fα(ω, t), Gα(ω, t))gF (α)gG(α)dαdα
by using the bilinearity of v(·, ·) and passing to the limit in the Riemann integral
over R2, which proves our assertion. The extension to all F,G ∈ S(Ω) is granted
by linearity. Theorem 5 and Proposition 7 show us that S(Ω) is dense in H(Ω, Q)
in the ‖ · ‖H norm. This is enough to extend the result to H(Ω, Q), under the ‖ · ‖Hnorm since we know that the unique solution to (3.3) is continuous in the ‖ · ‖Hnorm. This concludes the proof. ¤
Definition 14. Let q(·, ·) ∈ B(C1b (Ω)) a nonnegative symmetric continuous bilinear
form on C1b (Ω) and ξ(w) a random field on C1
b (Ω), measurable with respect to
(W,F , P ). Assume that for any G ∈ Ccyl(Ω) there exists a standard Brownian
motion β(t,w)t≥0 adapted to the filtration Ftt≥0 such that, if G(ω(·)) is written
16 ILIE GRIGORESCU
G(ω, t) = a(ω(·))g(ω(t)) as in (2.4), the process (G(ω, t), ξ(w))t≥t′ starting at
t = t′ from (a(ω(·))g(ω(t′)), ξ(w)) ∈ Ft′ (also called the associate marginal process
of G) satisfies the SPDE
(3.4)
d(G(ω, t), ξ(w)) =σ2
2(∂2G(ω, t), ξ(w))dt + σ
√q(∂G(ω, t), ∂G(ω, t))dβ(t, w) .
Then we shall say that ξ satisfies the inductive SPDE (3.4) with correlation q(·.·)with respect to Q.
Remark 1: For any time t the functional of ω(·) equal to G(ω, t) = a(ω(·))g(ω(t))
belongs to Ccyl(Ω) if a(·), g(·) belong to the continuous class (respectively Ecyl(Ω)
if a(·), g(·) belong to the exponential class) so (G(ω, t), ξ(w)) is well defined.
Remark 2: For proving our main result we only need q(·, ·) defined in (4.2).
However, Theorem 4 is true for a general C1b (Ω) - valued bilinear form q(·, ·).
Theorem 4. Let q(·, ·) ∈ B(C1b (Ω)) be a positive symmetric bilinear form con-
tinuous with respect to the H(Ω, Q) norm. Assume also that q(·, ·) ∈ B(C1b (Ω)),
a positive symmetric bilinear form continuous with respect to the H(Ω, Q) norm,
solves inductively the equation (3.2) for q(·, ·) with initial marginal q0(·, ·) as in
Definition 13. If Z = Z(w) is a random field on C1b (Ω), measurable with respect to
F , such that
(i) the one-dimensional restriction Z0 of Z at time t = 0 is a centered Gaussian
with covariance q0(f − 〈f〉µ, g − 〈g〉µ) and
(ii) Z satisfies the inductive SPDE (3.4) for all G ∈ E(Ω),
then Z can be uniquely extended to a centered Gaussian random field on H(Ω, Q)
with covariance
σ2Z(F, G) = q(F − 〈F 〉Q, G− 〈G〉Q) ,
for any two F,G ∈ H(Ω, Q).
Remark 1: The condition that Z be a random field on C1b (Ω) is weaker than Z
be indexed by the Banach space H(Ω, Q). However, the iterated SPDE, Theorem
5 and the fact that q(·, ·) can be extended to B(H(Ω, Q)) will show that Z is a
random field on H(Ω, Q).
Remark 2: This is a uniqueness theorem only. For the covariance prescribed in our
problem (2.11) - (2.12) the existence is proven by direct verification of the conditions
of Theorem 4 applied to q(·, ·) from (2.12) with q(·, ·) given in (4.2). We note that
in the proof of Proposition 4 the bilinear form q(·, ·) is split in two components
AN INFINITE DIMENSIONAL CENTRAL LIMIT THEOREM 17
(4.3) and (4.4) corresponding to the “independent” and “coupled” parts of the
covariance.
Proof of Theorem 4: Let’s denote the value of (G(ω, t), Z(w)) by Z(G(ω, t)) for
cylinder functions G ∈ (E ∪S)(Ω). In general, we shall suppress w in the following.
We shall apply the Ito formula to the function φ(Z(G(ω, t)), with φ(x) = x2 and
obtain, after taking the expected value, that
d E[Z(G(ω, t))2] =(σ2E[Z(G(ω, t))Z(∂2G(ω, t))] + σq(∂G(ω, t), ∂G(ω, t))
)dt .
This proves that E[Z(G(ω, t))Z(F (ω, t))] satisfies (3.2) by polarization. Next, the
uniqueness argument of Proposition 3 implies that
E[Z(F (ω, t))Z(G(ω, t))] = q(F − 〈F 〉, G− 〈G〉) ≤ C‖F‖H‖G‖H .
We need to show that Z(G(ω, t)) is a centered Gaussian. For all G ∈ S(Ω)
one can write G(ω(·)) =∑
l Gl(ω(·)), where we made the assumption that the
summation runs through a finite set, with Gl ∈ Scyl(Ω). This implies that G(ω, t) =∫R
∑l[(al(ω) exp (iαω(t)))gl(α)]dα exactly as detailed in the proof of Proposition
3. Then, equation (3.4) gives the SPDE satisfied by the process Z(G(ω, t))t≥t′
starting at t = t′ from Z(a(ω(·))g(ω(t′))) (also called the associate marginal process
of G)
(3.5) dZ(G(ω, t)) =σ2
2Z(∂2G(ω, t))dt + σ
√q(∂G(ω, t), ∂G(ω, t))dβ(t) .
The inductive SPDE (3.4) for a function of type al(ω) exp (iαω(t)) reduces to a
classical Ornstein-Uhlenbeck process. We know by the induction hypothesis that
at start t = t′ the process was a centered Gaussian. Let’s denote by Z(Gα,l(ω, t))
the solution to (3.4) for Gα,l(ω, t) = al(ω) exp (iαω(t)). The first two terms of the
equation are linear. The quadratic form q(·, ·) has the property
q(∂Gα,l(ω, t), ∂Gα,l(ω, t)) = |α|2q(Gα,l(ω, t), Gα,l(ω, t)) .
We derive that
Z(Gα,l(ω, t))− Z(Gα,l(ω, t′)) +α2σ2
2
∫ t
t′Z(Gα,l(ω, s))ds−
−|α|σ∫ t
t′
√q(Gα,l(s), Gα,l(s))dβ(s) = 0 .
18 ILIE GRIGORESCU
The initial Z(G(ω, t′)) is a mean zero normal random variable by the induction
hypothesis. We integrate the Ornstein-Uhlenbeck SDE t → Z(G(ω, t)) and obtain
(3.6) Z(G(ω, t)) =∫
R
∑
l
Z(Gα,l(ω, t))gl(α)dα
=∫
R
∑
l
Z(Gα,l(ω, t′)) exp (−α2σ2
2(t− t′))gl(α)dα
+σ
∫ t
t′
∫
R
∑
l
exp (−α2σ2
2(t− s))(
√q(Gα,l(s), Gα,l(s)))|α|gl(α)dαdβ(s) .
We need to justify the integration along the real line of the solutions Z(Gα,l(ω, t)).
Let G ∈ Ccyl(Ω), G(ω, t) = a(ω)g(ω(t)) such that a(ω) ∈ C(Ω), g ∈ S(X) and, for
any r ∈ Z+, let
(3.7) Rr(G(ω, t)) = a(ω)∑
−r2≤k≤r2
eiαkω(t)g(αk)(αk+1 − αk)
be the Riemann sum for the partition ∆r obtained by dividing the interval [−r, r]
into 2r2 equal subintervals with partition points denoted by αk, −r2 ≤ k ≤ r2.
Then, Rr(G(ω, t)) ∈ C(Ω) and
limr→∞
‖Rr(G(ω, t))−G(ω, t)‖H = 0 .
This follows from the properties of the inverse Fourier transform on the real line,
since a(ω) is a bounded smooth functional.
The real integral
(3.8) G(ω, t) =∫
R
( ∑
l
(al(ω) exp (iαω(t)))gl(α))dα
is the ‖·‖H - limit of Riemannian sums Rm(∑
l Gl(ω, t)) (equation 3.7). Since there
exists a constant C such that E[Z(G)2] ≤ C‖G‖2H, we can derive that the sequence
Z(∑
l Rm(Gl(ω, t)))m≥1 is tight with limit Z(G(ω, t)). A similar reasoning proves
that the last integral involving the bilinear form q(·, ·) can be integrated along α ∈ Ras a consequence of Plancherel’s identity for g′(x).
We notice that if
V (ω, t′) = E[∑
l
al(ω)gl(ω(t)) | Ft′]
then∫
R
∑
l
Z(Gα,l(ω, t′)) exp (−α2σ2
2(t− t′))gl(α)dα = 2πZ(V (ω, t′))
AN INFINITE DIMENSIONAL CENTRAL LIMIT THEOREM 19
which implies from the induction hypothesis that the first term in (3.6) is a zero
mean Gaussian measurable with respect to Ft′ . V (ω, t′) is indeed in S(Ω) as the
summation of convolutions of functions in S(Ω) with the transition probability of
Q from time t′ to time t. The second term is a stochastic integral of a deterministic
function of time depending on G against a Brownian motion on [t′, t], that is a
process with independent increments. The sigma fields they are supported on are
independent. This concludes the induction step needed to prove that the solution
is a mean zero normal random variable.
By piecing together the time marginals present in any G ∈ S(Ω) we have shown
that there exists a unique mean zero Gaussian random field satisfying the inductive
SPDE (3.4) for any G ∈ S(Ω) provided we start at time t = 0 with a random
field compatible with q(·, ·), that is, having the one dimensional bilinear form equal
to the marginal bilinear form of q(F − 〈F 〉µ, G − 〈G〉µ) at t = 0 (Definition 11).
It has been shown in the first part of the proof that the covariance is exactly
q(F − 〈G〉Q, F − 〈G〉Q). This and the continuity of q(·, ·) with respect to the Hnorm enables us to extend Z to H(Ω, Q), which concludes the proof. ¤
Corollary 1. Let q(·, ·), q(·, ·) and q0(f−〈f〉µ, g−〈g〉µ) be exactly like in Theorem
4. If there exists a random field Z on C1b (Ω) such that the restriction Z0 of Z to
one-dimensional functionals of the type (3.1) is a centered Gaussian with covariance
q0(f − 〈f〉µ, g − 〈g〉µ) and for any F,G ∈ E(Ω) the processes
Z(G(ω, t))− Z(G(ω, t′))− σ2
2
∫ t
t′Z(∂2G(ω, s))ds
and
Z(F (ω, t))Z(G(ω, t))− Z(F (ω, t′))Z(G(ω, t′))−σ2
2
∫ t
t′
(Z(∂2F (ω, s))Z(G(ω, s)) + Z(F (ω, s))Z(∂2G(ω, s))+
2q(∂F (ω, s), ∂G(ω, s)))ds
are (Ft, P ) - martingales, then Z is unique and has covariance q(F − 〈F 〉Q, G −〈G〉Q).
Remark: It is easy to see that G(ω) = πt(ω) = ω(t) belongs to H(Ω, Q) for any
time t ≥ 0. The corollary implies that Z(πt(ω)) is a Brownian motion with respect
to the filtration of the process.
20 ILIE GRIGORESCU
4. The proof for a special class of functions
The proof of Theorem 2 is the main result of this section.
Proposition 4. The bilinear form q(·, ·) ∈ B(H(Ω, Q)), defined in (2.12) as
q(F, G) =∫
Ω
(F (ω)− 〈F 〉Q)(G(ω)− 〈G〉Q)dQ(ω)+
+γ
∫
Ω
〈∇ηF 〉Q〈∇ηG〉QdQ0(η)
is symmetric, nonnegative, translation invariant and satisfies the inductive equation
(3.2) starting at q0(f, g) = Cov(f, g) for any f, g ∈ C1b (X), where
(4.1) Cov(f, g) =∫
X
(f(x)− 〈f〉µ)(g(x)− 〈g〉µ) µ(dx)
with correlation
(4.2) q(F, G) =∫
Ω
F (ω)G(ω)dQ(ω) + γ
∫
Ω
〈∇ηF 〉Q〈∇ηG〉QdQ0(η) .
Proof: Let
(4.3) q1(F,G) = EQ[(F (ω)− 〈F 〉)(G(ω)− 〈G〉)
]
and
(4.4) q2(F, G) = EQ0[(〈∇ηF 〉Q)(〈∇ηG〉Q)
].
We want to show that both satisfy the inductive equation (3.2) with
q1(∂F, ∂G) = EQ[(∂F (ω))(∂G(ω))]
and
q2(∂F, ∂G) = (EQ[∂F (ω)])(EQ[∂G(ω)]) .
The equation for q1(∂F, ∂G) is obtained as in the case of uncorrelated martingales
by Ito formula. We look at q2(·, ·) only. Let G(ω, t) = aG(ω)gG(ω(t)) and F (ω, t) =
aF (ω)gF (ω(t)). Then, if 〈·〉ω denotes the expected value with respect to Q,
(4.5) 〈∇ηG〉Q = 〈∇ηaG(ω)gG(ω(t))〉ω + 〈aG(ω)g′G(ω(t))〉ωη(t) ,
(4.6)ddt 〈∇ηaG(ω)gG(ω(t))〉ω = σ2
2 〈∇ηaG(ω)g′′G(ω(t))〉ωddt 〈aG(ω)g′G(ω(t))〉ω = σ2
2 〈aG(ω)g′′′G (ω(t))〉ω.
AN INFINITE DIMENSIONAL CENTRAL LIMIT THEOREM 21
The analogous formulas hold for F . The dt term from the quadratic variation of
the Brownian motion η(·) is
σ2(〈aF (ω)g′F (ω(t))〉ω〈aG(ω)g′G(ω(t))〉ω
)dt ,
equal to σ2q2(∂F (ω, t), ∂G(ω, t))dt. We superpose the two solutions and obtain
(3.2) for our random field ξ. ¤Remark. An alternate way to prove the proposition is to consider a pair of i.i.d.
Brownian motions ω1(·), ω2(·) with law Q = Qµ and another independent Brownian
motion η(·) with law Q0, write the Ito formula for (4.3) - (4.4) (before averaging)
for the three dimensional system, and finally take the expected value.
Proposition 5. For any G ∈ (E ∪ S)(Ω), that is such that there exists a positive
integer m and a function g(x0, x1, . . . , xm) in the linear span of cylinder functions
obtained as products of either S(R) or E(R) for which G(ω) = g(ω(t0), . . . , ω(tm)),
(4.7) lim supN→∞
E|(G, ξN )|2 ≤ c(G)
with c(G) depending only on g and T .
Proof of Proposition 5: We shall proceed by induction to show (4.7). If m = 1
and t0 = 0 we have the classical central limit theorem. Assume (4.7) is valid for
G depending on at most m − 1 times. In general we only need to do the proof
for cylinder functions since the extension to the general class of functions needed
in the proposition is done by linearity and depends, as required, only on g and
T . To make things precise g(x0, x1, . . . , xm) = Πmi=0gi(xi) and the constant c(g, T )
corresponding to Πm−1i=0 gi(xi) is denoted by c(g, m− 1, T ). Also we denote tm−1 by
t′. Even though we need g(x) in the Schwartz class, we shall investigate first the
case g(x) = exp (iαx) where i =√−1. We shall use the notation G(ω, t) = Gα(ω, t)
for such a test function later on in the proof when needed.
For any F,G ∈ C(Ω) we can write the generalized Ito formula for stochastic integrals
with respect to martingales (see [9]) for the product (F (ω, t), ξN )(G(ω, t), ξN )t≥t′ .
We recall that for every finite N the random fields ξN are finite random measures
on the path space Ω and
∂
∂t〈G(ω, t)〉Q =
σ2
2〈∂2G(ω, t)〉Q .
Then
(4.8) (F (ω, t), ξN )(G(ω, t), ξN )− (F (ω, t′), ξN )(G(ω, t′), ξN ) =
22 ILIE GRIGORESCU
12N
∫ t
t′
∑
1≤i,j≤N
[(σ2N,i(s)∂
2F (zi(·), s)− σ2〈∂2F 〉Q)(G(zi(·), s)− 〈G〉Q)
+(F (zi(·), s)− 〈F 〉Q)(σ2N,j(s)∂
2G(zj(·), s)− σ2〈∂2G〉Q)]
2∫ t
t′[1N
∑
1≤i≤N
(∂F (zi(·), s))(∂G(zi(·), s))(σ2N,i(s))]ds
[1
N2
∑
1≤i,j≤N
(∂F (zi(·), s))(∂G(zj(·), s))(NkijN (s))]ds + MN (t)
where MN (t) is a martingale. We can re-write the above formula as
(F (ω, t), ξN )(G(ω, t), ξN )− (F (ω, t′), ξN )(G(ω, t′), ξN ) =
σ2
21N
∫ t
t′
( ∑
1≤i,j≤N
[(∂2F (zi(·), s)− 〈∂2F 〉Q)(G(zj(·), s)− 〈G〉Q)
+(∂2G(zj(·), s)− 〈∂2G〉Q)(F (zi(·), s)− 〈F 〉Q)]
2[1N
∑
1≤i≤N
(∂F (zi(·), s))(∂G(zi(·), s))])ds
+σ2γ
∫ t
t′[
1N2
∑
1≤i,j≤N
(∂F (zi(·), s))(∂G(zj(·), s))]ds + MN (t) + EN (t) .
The error term is
(4.9) |EN (t)| ≤ C(F,G)1N
N∑
i=1
∫ t
t′(√
N |σ2N,i(s)− σ2|)ds+
+1
N2
∑
1≤i,j≤N
∫ t
t′|Nki,j
N (s)− γσ2|ds ,
where C(F, G) is a constant depending on the supremum over Ω of the functions F ,
∂F , ∂2F and the analogue values for G. More precisely EN (t) =∫ t
t′ eN (s)ds with
(4.10) lim supN→∞
E[E2(t)] ≤ TE[ ∫ T
0
e2N (s)ds
]
by Schwarz inequality for the time integral. The latter bound goes to zero as
N → ∞ from formula (4.9) plus Assumption 1, equations (2.9)-(2.10), and once
more by Schwarz applied to the average over N .
Let F = G = a(ω)g(ω(t)) with g(x) = exp (iαx). If we denote uN (t) =
E|(G(ω, t), ξN )|2 we derive
(4.11) uN (t)− uN (t′) = −α2σ2
∫ t
t′uN (s)− q(G(ω, s), G(ω, s))ds + EN (t)
where we denote, for simplicity,
q(s) = q(G(ω, s), G(ω, s)) =∫
Ω
G(ω, s)G(ω, s)dQ(ω)+
AN INFINITE DIMENSIONAL CENTRAL LIMIT THEOREM 23
+γ
(∫
Ω
G(ω, s)dQ(ω)∫
Ω
G(ω, s)dQ(ω))≤ 2( sup
x∈Rm+1|g(x)|)2 .
We differentiate, solve the ODE and obtain
(4.12) uN (t) = uN (t′) exp (−α2σ2(t− t′)) + α2σ2
∫ t
t′q(s) exp (−α2σ2(t− s))ds
+∫ t
t′eN (s) exp (−α2σ2(t− s))ds .
This shows that
(4.13) lim supN→∞
uN (t) ≤ lim supN→∞
uN (t′)+α2σ2Tcm(g) = c(g, m−1, T )+α2σ2Tcm(g)
where we denoted 2(supx∈Rm+1 |g(x)|)2 by cm(g).
Since lim supN→∞ uN (t′) ≤ c(g,m − 1, T ) by the induction hypothesis, relation
(4.12) proves the tightness of (G(ω, t), ξN )N in the special case g(x) = exp (iαx).
Let g(x) be a function in the Schwartz space and let g(x) =∫R g(α) exp (iαx)dα
where g(α) ∈ S(C) as well. Before passing to the limit as N → ∞ the random
fields ξN are finite random measures on Ω. The integration over the real line of α
can be viewed as the limit of finite Riemann sums converging in the uniform norm
on Cb(Ω), since, for a finite N , the random fields ξN are finite measures on Ω. We
can write
(G(ω, t), ξN ) =∫
R(Gα(ω, t), ξN )g(α)dα .
The sequence will be tight if lim supN→∞E|(G(ω, t), ξN )|2 < ∞. We calculate
E|(G(ω, t), ξN )|2 = E|∫
R(Gα(ω, t), ξN )g(α)dα|2 ≤
≤∫
RE|(Gα(ω, t), ξN )|2g2(α)dα .
Then, according to equation (4.13),
lim supN→∞
E|(G(ω, t), ξN )|2 ≤
≤∫
R
(c(g, m− 1, T ) + α2σ2Tcm(g)
)g2(α)dα ≤ c(g, m, T ) .
The very last step is to consider a general function G which will be a finite sum of
cylinder functions. The bound we obtain will depend on the number of terms in
G, determined exclusively by the test function g and the endpoint T of the time
interval. ¤
24 ILIE GRIGORESCU
Proposition 6. For any pair F, G ∈ (E ∪ S)(Ω) as in Proposition 5, the fami-
lies of processes (F (ω, ·), ξN )N>0 and (G(ω, ·), ξN )N>0 defined for t ≥ t′ (as
in formula (2.4) ) are tight and any pair of limit points, denoted by Z(F (ω, ·))and Z(G(ω, ·)) respectively, satisfy the inductive SPDE (3.4) with correlation form
(4.2).
Proof of Proposition 6: The m = 0 and t = 0 case is the classic central limit theorem.
At time t = t′ the functionals (G(ω, t′), ξN ) are uniformly square integrable in N
either by the induction hypothesis or directly from Proposition 5. The differences
(G(ω, t), ξN )− (G(ω, s), ξN ) will be treated analogously with (4.8).
(4.14) (G(ω, t), ξN )− (G(ω, s), ξN ) =
12√
N
∫ t
s
∑
1≤j≤N
(σ2N,j(u)∂2G(zj(·), u)− σ2〈∂2G(ω, u)〉Q)
+1√N
∫ t
s
∑
1≤j≤N
∂G(zj(·), u)dzj(u)
which can be written as
(4.15) (G(ω, t), ξN )− (G(ω, s), ξN ) =
σ2
2√
N
∫ t
s
∑
1≤j≤N
(∂2G(zj(·), u)− 〈∂2G(ω, u)〉Q)+
1√N
∫ t
s
∑
1≤j≤N
∂2G(zj(·), u)dzj(u) +∫ t
s
eN (G, u)du
with error term less than (4.9). We compute the square of the expectation of the
difference. The right hand side terms behave as follows.
lim supN→∞
E
∣∣∣∣∣∣1√N
∫ t
s
∑
1≤j≤N
(∂2G(zj(·), u)− 〈∂2G(ω, u)〉Q)du
∣∣∣∣∣∣
2
≤
∫ t
s
lim supN→∞
E∣∣(∂2G(ω, u), ξN )
∣∣2 du ≤ (t− s)C(∂2G)
according to Proposition 5 (we recall that C(∂2G) did not depend on any particular
time t). The martingale term will satisfy a similar inequality due to Lemma 3,
equation (6.2). We use the fact that
(4.16) 1√
N
∫ t
s
∑
1≤j≤N
∂G(zj(·), u)dzj(u)
2
− 1N
∫ t
s
∑
1≤j≤N
(∂G(zj(·), u))2σ2N,i(u)du−
AN INFINITE DIMENSIONAL CENTRAL LIMIT THEOREM 25
− 1N2
∫ t
s
∑
1≤i 6=j≤N
(∂G(zi(·), u))(∂G(zj(·), u))(NkNi,j)(u)du
is a martingale. We can substitute σ2N,i(u) with σ2 and NkN
i,j(u) by γσ2 due to
(4.9) and (4.10). In the limit as N →∞, the law of large numbers (Proposition 5)
yields a quadratic variation equal to exactly σ2ξ (∂F, ∂G).
We have shown, based on Proposition 5, that for any F, G ∈ C(Ω) we can con-
sider a weak limit of (F, ξN ) and (G, ξN ) denoted by Z(F ) and Z(G), respectively.
Formulas (4.15) and (4.16) imply that Z(G(ω, t)) satisfies the inductive SPDE:
(4.17) Z(G(ω, t))− Z(G(ω, t′)) =σ2
2
∫ t
t′Z(∂2G(ω, s))ds+
+σ
∫ t
t′
√q(∂G(s), ∂G(s))dβ(s)
with q(F, G) defined in (4.2). The calculations involved in determining the covari-
ance are direct consequences of Proposition 4. The analogous formula holds for
F and the product u(F,G) = Z(F (ω, t))Z(G(ω, t)) satisfies the inductive property
that
(4.18)
u(F (ω, t), G(ω, t))− u(F (ω, t′), G(ω, t′))−σ2
2
∫ t
t′
(u(∂2F (ω, s), G(ω, s)) + u(F (ω, s), ∂2G(ω, s))
+2q(∂F (ω, s), ∂G(ω, s))
)ds
is an (P,Ft)-martingale. Taking the expected value, equation (4.18) becomes (3.2).
¤
Proof of Theorem 2. Proposition 5 proves that the fluctuation random fields
ξNN>0 are tight in the weak* topology over the space of special functions (E ∪S)(Ω). Proposition 6 proves the conclusion of the theorem for functions in the
special class (E ∪ S)(Ω) based on Theorem 4 in Section 3 after identifying the
covariance function from Proposition 4. ¤
5. Extension to C1b (Ω)
This section proves Theorem 3 via Proposition 10. Suppose we can figure out
the covariance of the limiting random field in the central limit theorem for cylinder
functions. If the covariance is continuous in some norm on C1b (Ω), then we can define
directly the limit as a Gaussian on the completion of the new space. Identifying
the limit in Theorem 2 does not provide us with a class of functions for which the
26 ILIE GRIGORESCU
central limit theorem takes place, except for Schwartz or exponential class cylinder
functions, which may be a rather poor space. However, Proposition 9, together
with the density result from the next theorem overcome this difficulty and enable
us to prove Theorem 3 in an infinite dimensional setting using the test function
space C1b (Ω).
The proof of Proposition 9 requires that we evaluate the L2 norm of the renormal-
ized differences (amplified by a factor of√
N) between the martingales zi(·)1≤i≤N
and the limiting Brownian motions distributed according to Q from the decompo-
sition Lemma 2, which is the missing link between Theorem 5 and Proposition 9
- see also the comment related to (2.15). It seems hard to connect the supremum
norm of the path space Ω and the expectation of its square except by assuming
that the error terms are essentially martingales with quadratic variation of O( 1N )
and using Doob’s inequality with an exponential bound guaranteed by Proposition
8. Proposition 10, which finishes the proof of Theorem 3 is based on Proposition 9.
Theorem 5. Let G ∈ C1b (Ω). If K is a compact with K ⊆ Ω, then for any ε > 0
there exists a function GSε ∈ S(Ω) such that ‖GS
ε ‖C1b≤ 2‖G‖C1
band
(5.1) supω∈K
(|G(ω)−GS
ε (ω)|+ ‖∇·G(ω)−∇·GSε (ω)‖)
< ε
The proof of Theorem 5 is in the Appendix.
Proposition 7. The space S(Ω) is dense in H(Ω, Q).
Proof of Proposition 7: Q is a Brownian motion hence it is supported on the
countable union of compacts of the form Kα = Ω0 ∩ B(0, α), where B(0, α) is the
ball of radius α centered at ω = 0 and α > 0 converges to ∞ meanwhile Ω0 is
the set of Holder continuous paths ω ∈ Ω with exponent ν ∈ (0, 12 ). These sets
are equicontinuous and bounded in the supremum norm, which implies they are
compact by Arzela’s theorem. The functions G ∈ C1b (Ω) have finite ‖ · ‖H norm.
For every G the measures (|G(ω)|2 + ‖∇·G(ω)‖2)dQ(ω) are absolutely continuous
with respect to Q and∫
Ω0
|G(ω)|2 + ‖∇·G(ω)‖2dQ(ω) =∫
Ω
|G(ω)|2 + ‖∇·G(ω)‖2dQ(ω) ,
hence
limα→∞
∫
Kcα
|G(ω)|2 + ‖∇·G(ω)‖2dQ(ω) = 0
AN INFINITE DIMENSIONAL CENTRAL LIMIT THEOREM 27
by dominated convergence. This implies that there is a compact Kα0 with the
property∫
Kcα0
dQ(ω) < ε and∫
Kcα0
|G(ω)|2 + ‖∇·G(ω)‖2dQ(ω) < ε .
Finally, for this compact we pick a GSε ∈ S(Ω) such that (5.1) and ‖GS
ε ‖C1b≤
2‖G‖C1b
are satisfied. This and the Schwartz inequality imply that
‖G−GSε ‖2H ≤
∫
Kα0
[|G(ω)−GS
ε (ω)|2 + ‖∇·(G(ω)−GSε (ω))‖2
]dQ(ω)
+2ε + 2‖GSε ‖2C1
b
∫
Kcα0
dQ(ω) ≤ (3 + 8‖G‖2C1b)ε .
The claim is proven. ¤Lemma 1 is an independent result valid in any inner product space. It is not
essential that the space be complete (hence Hilbert). We shall use the standard
notation ‖ · ‖ for the norm. This should not be confused with the supremum norm
on Ω the path space of continuous functions. The heuristic argument relating it to
our proof is that the inner product is analogous to the quadratic variation.
Lemma 1. Let X be an inner product space with inner product denoted by (·, ·).The associated norm will be denoted by ‖ · ‖. For any positive integer N ∈ Z+ we
shall consider a family of vectors vNi 1≤i≤N with the property that there exist two
numbers σ > 0 and γ > −1, independent of N such that
(5.2) limN→∞
√N max
1≤i≤N|(vN
i , vNi )− σ2| = 0
and
(5.3) limN→∞
max1≤i,j≤N
|N(vNi , vN
j )− γσ2| = 0 .
Then, for N sufficiently large, the matrix V N with elements V Nij = (vN
i , vNj ) is
positive definite and there exists an orthonormal system of N vectors wNi 1≤i≤N
and a constant C, independent of N , such that
max1≤i≤N
‖vNi − σwN
i ‖2 ≤C
N.
Proof: the proof of Lemma 1 is in the Appendix.
28 ILIE GRIGORESCU
Lemma 2. Assume that conditions (2.13) and (2.14) are met. Then, for any N ∈Z+ there exist N (P,Ft)- martingales yN
i (·)1≤i≤N such that σ−1yNi (·)1≤i≤N
is a standard N -dimensional Brownian motion with the property that the renor-
malized differences dNi (·) =
√N(zN
i (·) − yNi (·)) are continuous square-integrable
martingales with quadratic variations 〈dNi , dN
i 〉(·) for which there exists a constant
C, independent of N , the time t ∈ [0, T ] and the indices 1 ≤ i ≤ N , such that
(5.4) 0 ≤ d
dt〈dN
i , dNi 〉(t) ≤ C
P -almost surely.
Remark: The martingale representation theorem implies that the quadratic vari-
ation process of a square-integrable martingale relative to the filtration Ftt≥0 is
absolutely continuous P -almost surely with respect to the Lebesgue measure.
Proof of Lemma 2: The martingale representation theorem (in [9], page 84) and
the fact that the covariance matrix of the martingales zNi (·)1≤i≤N is positive
definite (Lemma 1) ensure the existence of a system of N independent Brownian
motions β1(·), . . . , βN (·) on (W, P,F), adapted to the original filtration Ftt≥0,
and a set of progressively measurable square integrable system of functions ψNil (t),
with 1 ≤ l ≤ N and 1 ≤ i ≤ N such that, P -almost surely,
zNi (t) = zN
i (0) +N∑
l=1
∫ t
0
ψNil (s)dβl(s) .
Let HN be the subspace generated by the martingales βi(·)1≤i≤N in L2(Ω, P ),
that is, the completion under the L2(W, P ) norm of the martingales of the form
(5.5) M(t) = M(0) +N∑
l=1
∫ t
0
rl(s)dβl(s)
with rl(s) , 1 ≤ l ≤ N a family of bounded progressively measurable functions with
respect to Ftt≥0. Let M(·) be an element from HN . M(t) can be written in
terms of square integrable functions rl(s) as in (5.5). For every u ∈ [0, T ] we shall
define the mapping Su from HN to RN by
Su(M(·)) = (r1(u), . . . , rN (u)) .
If M1(·) and M2(·) are two elements of HN , the quadratic variation will be the
time integral of the Euclidian inner product on RN for the vectors Su(M1(·))
AN INFINITE DIMENSIONAL CENTRAL LIMIT THEOREM 29
and Su(M2(·)). The martingale representation theorem implies that any square-
integrable martingale will have an absolutely continuous quadratic variation P -
almost surely. As a consequence, we can evaluate the derivative of the quadratic
variation process for any martingale in HN directly from the Euclidian norm of the
vector Su(M(·)). Let N− 12 dN
i (·), for 1 ≤ i ≤ N be the differences between the
original martingales and the orthonormal set of martingales produced by Lemma
1. All the transformations involved in the procedure are progressively measurable.
We always can start our orthonormal set of martingales and the original set of
martingales from the same points so that the differences actually start at zero.
Assumptions (2.13) and (2.14) imply that the conditions of Lemma 1 are satisfied
with nonrandom bounds independent from N and the indices 1 ≤ i ≤ N for all
Su(zNi (·)). This implies that the difference processes dN
i (·)1≤i≤N will satisfy
the property
(5.6) 0 ≤ d
du〈dN
i , dNi 〉(u) ≤ C
P -almost surely for some universal constant C. This proves the lemma. ¤
Proposition 8. Let mN (·)N be a family of continuous square integrable martin-
gales with respect to (P, Ft≥0t≥0) starting from a sequence of uniformly bounded
values mN (0). For any N ∈ Z+ there exist random functions aN (t) and a constant
K > 0 independent from N , such that the cross-variation process of mN (·) satisfies
〈mN , mN 〉(t) =∫ t
0
aN (s)ds
with |aN (t)| ≤ K P -almost surely. Then, for any t > 0 and ` ∈ R,
lim supN→∞
E[exp (`|mN (t)|)] < ∞ .
Proof of Proposition 8: Since exp (|z|) ≤ exp (z) + exp (−z) it is enough to check
what happens for exp (`mN (t)) with ` ∈ R. We apply the Ito formula to the
function z −→ exp (`z) and obtain
exp (`mN (t))− exp (`mN (0))− `2
2
∫ t
0
exp (`mN (s))aN (s)ds
is a martingale with respect to the same filtration and P . The expected value shows
that if we denote wN (t) = E[exp (`mN (t))] then
wN (t) ≤ wN (0) +K`2
2
∫ t
0
wN (s)ds
30 ILIE GRIGORESCU
which, in its turn, proves that wN (t) ≤ wN (0) exp (K`2t2 ) for all 0 ≤ t ≤ T . Since
wN (0) and t are bounded quantities we are done. ¤
Remark: Proposition 8 is much stronger than what we need, that is the r -
integrability of mN (t) for some r ≥ 4.
Proposition 9. There exists a constant C depending only on T such that for any
G ∈ C1b (Ω)
(5.7) lim supN→∞
E[|(G, ξN )|2] ≤ C‖G‖2H .
Proof of Proposition 9: Let yNi (·), 1 ≤ i ≤ N be the independent Brownian motions
from Lemma 2.
1√N
N∑1
(G(zNi (·))− 〈G〉Q) =
1√N
N∑1
(G(zNi (·))−G(yN
i (·))
+1√N
N∑1
(G(yNi (·))− 〈G〉Q) .
Let’s denote dNi (·) =
√N(zN
i (·)− yNi (·)). We decompose the first term into
(I) =1√N
(N∑1
G(zNi (·))−G(yN
i (·))− 1√N∇GdN
i (·)(yNi (·))
)
and
(II) =1N
N∑1
∇dNi (·)G(yN
i (·)) .
Again, we look at the two terms in this formula separately. We recall (2.1) from
Definition 1 and see that we can write an upper bound for |(I)| as
1√N
N∑1
∣∣∣∣G(zNi (·))−G(yN
i (·))− 1√N∇GdN
i (·)(yNi (·))
∣∣∣∣ ≤
≤√
Nc(G,1√N
)
(1N
N∑1
‖dNi (·)‖2
).
This quantity goes to zero as N →∞ as soon as the martingales dNi (·) have a finite
second moment. This is guaranteed by Proposition 8 (the uniformity in N and j)
and Doob’s maximal inequality regarding any Lr norm of martingales (r > 1) for
all 1 ≤ i ≤ N :
(5.8) (E[| sup0≤t≤T
di(t)|r]) 1r ≤ r
r − 1(E[|di(T )|r]) 1
r .
AN INFINITE DIMENSIONAL CENTRAL LIMIT THEOREM 31
The term
E[|(II)|2] = E
[1N
N∑1
∇√N(zNi (·)−yN
i (·))G(yNi (·))
‖√N(zNi (·)− yN
i (·))‖ · ‖√
N(zNi (·)− yN
i (·))‖]2
can be estimated if we expand the square after taking the supremum norm for the
linear operators ∇·G(yNi (·)) (as in Definition 1)
(5.9) E
[1N
N∑1
‖∇·G(yi(·))‖‖dNi (·)‖
]2
≤ E[ 1N2
N∑1
(‖∇·G(yi(·))‖‖dNi (·)‖)2
]+
1N2
∑
1≤i 6=j≤N
E[(‖∇·G(yi(·))‖‖∇·G(yj(·))‖)(‖dN
i (·)‖‖dNj (·)‖)
].
The first term can be bounded by using the supremum c1(G) of all ‖∇·G(ω(·))‖(Definition 1). The bound is
c1(G)21
N2
∑
1≤i6=j≤N
E[(‖dNi (·)‖2)] = O(
1N
) .
The second term can be bounded in terms of the L2 norm of ‖∇·G(yi(·))‖ due to the
crucial fact that yi(·) and yj(·) are independent Brownian motions by construction.
The upper bound is obtained by Schwarz’s inequality
1N2
∑
1≤i 6=j≤N
(E[(‖∇·G(yi(·))‖2‖∇·G(yj(·))‖2)]) 12 (E[(‖dN
i (·)‖‖dNj (·)‖)2]) 1
2
≤ ‖G‖2H1
N2
∑
1≤i 6=j≤N
DiDj
where Di = (E[‖di(·)‖4]) 14 for all 1 ≤ i ≤ N . These norms are bounded. We use
again Doob’s inequality (5.8) and Proposition 8 to ensure uniformity in both N
and i. Finally
lim supN→∞
E
[1√N
N∑1
(G(yNi (·)− 〈G〉Q)
]2
≤ E[G2 − (E[G])2] ≤ 2‖G‖2H
due to the central limit theorem variance for independent random variables. The
last two bounds provide an upper bound of the form C · ‖G‖2H with the constant
independent from N and G. This concludes the proof of the proposition. ¤
Proposition 10. Let ξNN>0 be a family of random fields on C1b (Ω) defined in
(2.11) and q(·, ·) ∈ B(H(Ω, Q)) be the bilinear form (2.12). Then, for any pair
FS , GS ∈ S(Ω)
(i) (FS , ξN ) and (GS , ξN ) converge weakly to mean zero normal random variables
with covariance q(FS − 〈FS〉Q, GS − 〈GS〉Q)
32 ILIE GRIGORESCU
(ii) limN→∞E[(FS , ξN )(GS , ξN )] = q(FS − 〈FS〉Q, GS − 〈GS〉Q).
(iii) for any G ∈ C1b (Ω), the sequence (G, ξN )N>0 is tight, and if F, G ∈ C1
b (Ω)
then the random variables (F, ξN ), (G, ξN ) converge weakly to centered Gaussians
with covariance q(F −〈F 〉Q, G−〈G〉Q). As a consequence, the unique limit ξ of the
sequence ξNN>0 is a centered Gaussian random field on H(Ω, Q) with the same
covariance.
Proof of Proposition 10: Proposition 6 and Theorem 4 show that the sequence
(GS , ξN )N>0 converges weakly to a centered normal random variable Z(GS).
Moreover, Z(GS) has variance equal to q(GS − 〈GS〉Q, GS − 〈GS〉Q). This proves
(i) and (ii).
Let l ∈ Z+. Theorem 5 and Proposition 7 indicate that for any G ∈ C1b (Ω) we
can choose the appropriate GSε for ε = 1
l , which we shall denote by GSl , such that
(5.10) ‖G−GSl ‖2H ≤
1l
.
Equation (5.10) implies that liml→∞ ‖GSl −G‖H = 0, which also guarantees that GS
l
are uniformly square integrable, hence the sequence of centered Gaussian random
variables Z(GSl )l∈Z+ is tight. Let ZS(G) be a limit point. This has to be a
centered Gaussian (we can look at the characteristic function of the Gaussians)
and its variance will be (see [15])
(5.11) σ2(ZS(G)) = liml→∞
q(GSl − 〈GS
l 〉Q, GSl − 〈GS
l 〉Q) = q(G− 〈G〉Q, G− 〈G〉Q) .
We already know from (5.7) that (G, ξN )N>0 is tight as well. Let Z(G) be a
limit point of (G, ξN )N>0. We can restrict ourselves without loss of generality to
subsequences of l and N such that Z(GSl ) ⇒ ZS(G) and (G, ξN ) ⇒ Z(G).
For α ∈ R we consider
(5.12) |E[e−iαZS(G)]− E[e−iαZ(G)]| ≤
(5.13)
|E[e−iαZS(G)]− E[e−iαZ(GSl )]| + |E[e−iαZ(GS
l )]− E[e−iα(GSl ,ξN )]|
+ |E[e−iα(GSl ,ξN )]− E[e−iα(G,ξN )]| + |E[e−iα(G,ξN )]− E[e−iαZ(G)]| .
The third term in (5.13) has the upper bounds
|E[e−iα(GSl ,ξN )]− E[e−iα(G,ξN )]| ≤ E[|e−iα(G−GS
l ,ξN ) − 1|]
≤ 2E[sin( (G−GS
l , ξN )2
)] ≤ E[|(G−GS
l , ξN )|] ≤ E[|(G−GSl , ξN )|2] 1
2 .
AN INFINITE DIMENSIONAL CENTRAL LIMIT THEOREM 33
If we let N →∞ we obtain that (5.12) is less than
|E[e−iαZS(G)]− E[e−iαZ(GSl )]|+
√C‖G−GS
l ‖H
where C is the constant in (5.7). We obtain that E[e−iαZS(G)] = E[e−iαZ(G)] after
l →∞. The left hand term is the Fourier transform of a centered Gaussian, which
proves that the right hand side is a centered Gaussian as well for any limit point
of (G, ξN )N>0 and any G ∈ C1b (Ω). On the other hand, we have shown that Z(G)
has the same distribution as ZS(G), a weak limit of Gaussians Z(GSl ) as l → ∞,
which implies that the variance of the limit is the limit of the variances (5.11). We
can repeat this reasoning for G = c1G1 + c2G2 with arbitrary constants c1, c2 and
G1, G2 ∈ C1b (Ω) and conclude the proof by polarization. ¤
Proof of Theorem 3. We only have to apply Proposition 10 to the random fields
ξNN>0 from Theorem 2 from Section 4. ¤
6. Appendix
Proof of Theorem 1: In this case simple product functions are sufficient to prove
the theorem. We may choose them in S(Ω) without any loss of generality. We shall
proceed by induction on the number m of time marginals present in G. For m = 1
the martingale part vanishes as N → ∞ in Ito formula. The integrand of the dt
term is uniformly bounded by the supremum norm of ∂2G. The error terms have
bounds of order inferior to N−1, as prescribed by (4.9). This implies the tightness.
Any limit point will be deterministic, since the martingale part vanishes as N →∞and we can easily check that it must verify the same weak PDE (the heat equation)
as E[G(ω, t)]. The details of the proof of this type of result can be found, for
example, in [11] in Chapter 4, and also in [5]. The same reasoning applies when
we perform the induction step over m, since we are allowed to start over from an
arbitrary initial profile at time t′ = tm−1. We can pass to functions of Cb(Ω) class
due to the fact that
(6.1) lim supN→∞
E
∣∣∣∣∣1N
N∑
i=1
(G(zNi (·))− 〈G〉Q)
∣∣∣∣∣
2 ≤ 2(sup
ω∈Ω|G(ω)|)2 .
This concludes the proof. ¤
34 ILIE GRIGORESCU
Lemma 3. For any G ∈ C(Ω) as in (2.4)
(6.2) limN→∞
E
sup
t′≤t≤T
∣∣∣∣∣1N
N∑
i=1
(G(zN
i , t))− 〈G(ω, t)〉Q)∣∣∣∣∣
2 = 0 .
Proof of Lemma 3: Because of (6.1), it is sufficient to prove the lemma for cylinder
functions G ∈ Scyl(Ω). We write G(ω(·)) = g(ω(t0), . . . , ω(tm)) with g in the
Schwartz class on Rm+1 and let tm be denoted by t and tm−1 by t′ for simplification.
With the convention of notation from (2.4), we first establish (6.2) for a fixed
time t ∈ [t′, T ]. This is a consequence of the inductive proof from Theorem 1.
Since the test function is uniformly bounded and the quadratic variation of the
martingale part vanishes as N →∞ we can adapt the same argument by squaring
the differences and show (6.2) for every fixed time. In order to prove uniformity,
assume that there is a sequence of times such that the absolute value in (6.2) exceeds
a constant c > 0. Since the interval [0, T ] is compact, there is a subsequence of
times converging to some t′′ ∈ [0, T ]. However, the functionals are continuous in
time at t′′ and G(ω, t) approaches the value at t′′ according to the estimate
lim supN→∞
E
∣∣∣∣∣1N
N∑
i=1
(G(zN
i , t))− 〈G(ω, t)〉Q)− 1
N
N∑
i=1
(G(zN
i , t′′))− 〈G(ω, t′′)〉Q)∣∣∣∣∣
2
≤ 2 lim supN→∞
sup
x|∂G|2E
[1N
N∑
i=1
|zNi (t)− zN
i (t′′)|2]
+ 2Tσ2c1(G)|t− t′′|
≤ c2(G, T )σ2|t− t′′| .The error is independent of N and of the order of magnitude of t− t′′ . The error
obtained is arbitrarily small, a contradiction with the fact that the absolute value
in (6.2) exceeds a constant c > 0. This concludes the proof. ¤
Proof of Theorem 5: Let m ∈ Z+. The mapping Tm : Ω −→ Ω is defined
for each ω as the new continuous path obtained by linear interpolation between
the values ω(ti), for all 0 ≤ i ≤ m, at the points ti = iTm . It follows that Tm is
linear and continuous with ‖Tm‖ ≤ 1 in the supremum norm. Since Tm is linear
and continuous it is differentiable and ∇ηTmω = Tmη. We shall define the finite-
dimensional norm on ω ∈ Ω as
(6.3) ‖ω‖m = max0≤i≤m
|ω(ti)|
Any compact set in Ω is uniformly bounded and equicontinuous by Arzela -
Ascoli theorem as in [3]. We can assume that |ω(t)| ≤ M for all t ∈ [0, T ] and all
AN INFINITE DIMENSIONAL CENTRAL LIMIT THEOREM 35
ω ∈ K. The fact that G ∈ C1b (Ω) (Definition 1) and K is a compact in Ω implies
that for any ε > 0 there exists a δ = δ(ε) such that
|G(ω)−G(ω′)|+ ‖∇·G(ω)−∇·G(ω′)‖ < ε
if ω′(·) and ω(·) are in K and sup0≤t≤T |ω(t) − ω′(t)| < 3δ. We are free to choose
δ < min( ε2 , ε
‖G‖C1
b
). For any δ > 0 there exists an m ∈ Z+ such that
(6.4) |t− s| < 1m⇒ |ω(t)− ω(s)| < δ
uniformly in ω ∈ K.
For ε > 0 we shall choose a covering K ⊆ ∪ω∈KB(ω, δ) of the compact K with
balls of radius δ = δ(ε) in the uniform norm topology of Ω and extract a finite
subcovering with centers at ωj(·), j ∈ Jε, where Jε is a finite set depending only on
K and ε. For every x = (x0, x1, . . . , xm) ∈ Rm+1 we determine the path ωx as the
linear interpolation between the values (for 0 ≤ i ≤ m), that is ωx(ti) = xi. Let
Kxδ,j = x : ‖ωx − ωj‖m ≤ δ be the Rm+1 - cube of size 2δ and Kx
δ = ∪j∈JεKxδ,j .
We shall construct a function gε(x) on Rm+1 by piecing together the following
mappings. For m as in (6.4) and j ∈ Jε, let
(6.5) gm,j(x) = G(ωj) +∇Tm(ωx−ωj)G(ωj)
on each Rm+1 - cube Kxδ,j and zero everywhere else. Then, we define
(6.6) gε(x) = gm,j(x) if ‖ωx − ωj‖m = minj′∈Jε
‖ωx − ωj′‖m
with the understanding that if a point falls on the hypersurface where two or more
indices achieve the maximum we select the smaller index j′. This fact will not
affect the construction due to the mollification we do next. The function gε(x) is
piecewise smooth and uniformly bounded by
|gm,j(x)| ≤ supω∈Ω
|G(ω)|+ supω∈Ω
‖∇·G(ω)‖δ ≤ ‖G‖C1b.
The gradient of gε(x) can only be one of the linear mappings on Rm+1 from
the finite family of bounded linear operators ∇Tm(ωx−ωj)G(ωj). They are natu-
rally bounded in the supremum norm of continuous linear operators on Rm+1 by
supω∈Ω ‖∇·G(ω)‖. One can write that
(6.7) supx∈Rm+1
(|gε(x)|+ ‖∇gε(x)‖) ≤ supω∈Ω
|G(ω)|+ 2‖∇·G(ω)‖
36 ILIE GRIGORESCU
wherever the gradient is defined and note that supp(gε(x)) is a compact included
in Kxδ ⊆ [−M − 1,M + 1]m+1 in Rm+1.
For an arbitrary ρ > 0 we can construct a regularized version of gε(x) by convolu-
tion with φm,ρ(x) = ρ−m−1φ(ρ−1‖(x)‖Rm+1) where φ(x) = k0 exp ( 1x2−1 ) if |x| < 1
and identically zero outside the unit interval and ‖ · ‖Rm+1 is the Euclidian norm
on Rm+1. The constant k0 normalizes φ(x) so that the integral equals one. Let
(6.8) gSε (x) = (gε ∗ φm,ρ)(x) .
We choose ρ = ρ(ε) < δ2 to make sure that the function gε(x) vanishes outside the
compact Kx2δ ⊆ Rm+1. We shall use the observation that the convolution with φm,ρ
is a contraction in the following sense. If g(x) is a piecewise smooth function on
Rm+1 and gS(x) = (g ∗ φm,ρ)(x) then, for any pair of points x′ and x′′ from Rm+1,
(6.9) |gS(x′)− gS(x′′)| ≤ supy′∈B(x′,ρ),y′′∈B(x′′,ρ)
|g(y′)− g(y′′)|
and
(6.10) ‖∇gS(x′)−∇gS(x′′)‖ ≤ supy′∈B(x′,ρ),y′′∈B(x′′,ρ)
‖∇g(y′)−∇g(y′′)‖ .
This property equally allows us to estimate both
|gS(x)| ≤ supy∈B(x,ρ)
|g(y)| , ‖∇gS(x)‖ ≤ supy∈B(x,ρ)
‖∇g(y)‖ .
We define
(6.11) GSε (ω) = gS
ε (ω(t0), . . . , ω(tm)) .
The function gSε is of Schwartz class on Rm+1 but not of cylinder type. Any such
function can be approximated uniformly including its derivatives on any compact
by a linear combination of cylinder-type functions of Schwartz class. We shall show
this fact at the end of the proof. The norm ‖ω‖m determines a family of sets
Bm(ωj , δ) = ω : ‖ω − ωj‖m ≤ δ (not proper balls in Ω). We have to estimate
the differences
supω∈K
|GSε (ω)−G(ω)| ≤
≤ maxj∈Jδ
sup
ω∈Bm(ωj ,δ)∩K
|GSε (ω)− GS
ε (ωj)|+ |GSε (ωj)−G(ωj)|+
+ supω∈Bm(ωj ,δ)∩K
|G(ωj)−G(ω)|
and
supω∈K
‖∇·GSε (ω)−∇·G(ω)‖ ≤
AN INFINITE DIMENSIONAL CENTRAL LIMIT THEOREM 37
maxj∈Jδ
sup
ω∈Bm(ωj ,δ)∩K
‖∇·GSε (ω)−∇·G(ωj)‖+ ‖∇·GS
ε (ωj)−∇·G(ωj)‖+
+ supω∈Bm(ωj ,δ)∩K
‖∇·G(ωj)−∇·G(ω)‖
where we always choose to place ω in the ball Bm(ωj , δ) for which the center is
the closest element from ω among the finite collection of points ωj ∈ Jε. The
uniform continuity on K guarantees that the third terms on the right hand side
of both inequalities are of order ε. To estimate the differences |GSε (ωj) − G(ωj)|
and ‖∇·GSε (ωj) − ∇·G(ωj)‖ we need to look at the values of gε(x) and ∇gε(x)
in a neighborhood Kxδ,j′ , j′ ∈ Jε and compare to the values at ωj . The values
of gSε (ωj) are the result of convolution with φm,ρ of values of gε for some x with
‖ωx − ωj‖m < ρ. The worst case scenario is that the value of gε at x is a value of
gm,j′ on an adjacent ball Kxδ,j′ to Kx
δ,j (this fact grants that ‖ωj−ωj′‖m < 2δ < 3δ)
and then |GSε (ωj)−G(ωj)| ≤ |gm,j′(x)−G(ωj′)|+ |G(ωj′)−G(ωj)| which is of order
ε by construction. The same is valid for ‖∇·GSε (ωj)−∇·G(ωj)‖, bounded above by
supx∈Kx
δ,j′‖∇·gm,j′(x)−∇·G(ωj′)‖+ ‖∇·G(ωj′)−∇·G(ωj)‖ ,
of order ε as well.
We need to estimate the first terms on the right hand side of the inequalities.
|GSε (ω)−GS
ε (ωj)| and ‖∇·GSε (ω)−∇·GS
ε (ωj)‖ are bounded above by the supremum
value of the differences |gε(x′) − gε(x′′)| and ‖∇·gε(x′) − ∇·gε(x′′)‖, respectively,
where ‖ωx′ − ω‖m ≤ ρ and ‖ωx′′ − ωj‖m ≤ ρ. Assume the value at x′ is given by
gm,j′ on a ball Kxδ,j′ and the value at x′′ is given by gm,j′′ on a ball Kx
δ,j′′ . At this
point we intercalate the values of the functions at ωj′ , ωj and ωj′′ . The differences
between values in the same domain Kxδ,l, for any l ∈ Jε are of order ε. We only
have to compare the values at ωj′ and ωj′′ with ωj . The distance between ωj′′ and
ωj is less than δ + ρ. First,
‖ωj′′ − ωj‖m ≤ supx∈Kx
δ,j′′‖ωj′′ − ωx′′‖m + ‖ωx′′ − ωj‖m ≤ δ + ρ < 3δ
which implies that the error is of order ε. We know that ωj is the closest of all
ωl, with l ∈ Jε from ω. Hence ‖ω − ωj‖m ≤ ‖ω − ωj′‖m. In the same time
‖ω−ωx′‖m ≤ ρ and ‖ωx′ −ωj′‖m ≤ δ. We conclude that ‖ω−ωj′‖m ≤ δ +ρ which
implies that ‖ωj′−ωj‖m ≤ 2δ+2ρ < 3δ. The difference will be of order ε once again.
We obtained a function GSε (ω) = gS
ε (ω(t0), . . . , ω(tm)) where gSε (x) ∈ C∞0 (Rm+1)
within distance Cε from G inside the compact K in the uniform norm, where C is
38 ILIE GRIGORESCU
independent from G and K, and bounded by 2‖G‖C1b. The number m depends on
ε and the compact K.
We still have to prove that we can approximate GSε with a function G ∈ S(Ω). For
a given ε and a compact K, the function gSε (x) and m are fixed. The function gS
ε
has support included in the compact [−M−1,M +1]m+1 ⊆ (−M−3,M +3)m+1 ⊆Rm+1. Let K ′ = [−M − 2,M +2]m+1. It is known (for example from [10]) that for
any r ∈ Z+, a function f ∈ Cr(Rm+1) and all its derivatives can be approximated
uniformly on any compact with polynomials in Rm+1. The class of polynomials is of
cylinder type, in the sense that it is the linear span of products of functions (in this
case, polynomials of one variable) of the variables x0, x1, . . . , xm. The problem is
that these functions are not of Schwartz class. The indicator function of the compact
K ′ is the product of the indicator functions of the interval [−M − 2,M + 2], hence
of cylinder type. The product of the indicator functions with each polynomial will
be of cylinder type. We can consider the convolution with φm,ρ(x), with ρ < 12 .
Let ε′ be the accuracy of the approximation in the supremum norm. The function
φm,ρ(x) is of cylinder type as well. A consequence of this will be that the result
of the convolution will be of class C∞0 (Rm+1), will be of cylinder type (as the
convolution of two cylinder type functions, that is for which the variables decouple
in the convolution integral), and will stay within distance ε′ uniformly together
with all the derivatives.
We choose ε′ = εm+1 and let hε′(x) be the sum of cylinder functions with compact
support approximating gSε (x). The function GS
ε (ω) = hε′(ω(t0), . . . , ω(tm)) is in
S(Ω) and |GSε (ω)− GS
ε (ω)| ≤ ε′ < ε. We conclude the proof by noticing that
‖∇·GSε (ω)−∇·GS
ε (ω)‖ ≤∑
0≤i≤m
supx∈Rm+1
‖∂xihε′(x)−∇xigSε (x)‖ ≤ (m+1)ε′ = ε . ¤
Proof of Lemma 1: We can re-write the matrix V N = (V Nij ) as
(6.12) V Nii = σ2 +
cNi (σ)√
Nwith max
1≤i≤N|cN
i (σ)| ≤ cN (σ)
and
(6.13) V Nij =
γσ2
N+
cNij (γ)N
with max1≤i,j≤N
|cNij (γ)| ≤ cN (γ)
where limN→∞(cN (σ) + cN (γ)) = 0.
AN INFINITE DIMENSIONAL CENTRAL LIMIT THEOREM 39
We first prove the positive definiteness. Let x = (x1, x2, . . . , xN ) be an arbitrary
N dimensional vector. If S =∑
1≤i≤N xi
(x, V N x) =∑
1≤i,j≤N
V Nij xixj =
∑
1≤i≤N
(Viixi +
∑
j 6=i
Vijxj
)xi
=∑
1≤i≤N
(σ2 +
cNi (σ)√
N
)x2
i +∑
1≤i,j≤N,j 6=i
(γσ2
N+
cNij (γ)N
)xjxi =
(6.14) σ2( ∑
1≤i≤N
x2i +
γ
N(S − xi)xi
)+
(6.15)∑
1≤i≤N
cNi (σ)√
Nx2
i +∑
1≤i,j≤N,i 6=j
cNij (γ)N
xjxi .
We can bound (6.15) in absolute norm by
‖x‖2[(cN (σ)√
N+
cN (γ)N
)+ cN (γ)
( (∑
1≤i≤N |xi|)2N‖x‖2
)]
while this can be bounded above by Schwarz’s inequality
(6.16) ‖x‖2[cN (σ)√
N+ cN (γ)(1 +
1N
)]∼ o(1) .
The main term (6.14) is zero if ‖x‖ = 0, but otherwise it is equal to
σ2(1− γ
N
)‖x‖2 +
γσ2
NS2 = σ2‖x‖2
((1− γ
N) + γ
S2
N‖x‖2)
.
If γ ≥ 0 a lower bound is σ2‖x‖2(1 − γ
N
)of order O(1) and if γ < 0 we can
use Schwarz’s inequality again and obtain the lower bound σ2‖x‖2(1 + γ − γ
N
).
This proves that (x, V N x) ≥ σ2‖x‖2C ′ with C ′ > 0 independent of N . The only
condition needed to ensure a lower bound uniformly in N is γ > −1.
We proceed to the proof of the existence of wNi 1≤i≤N . We first write vN
i 1≤i≤N
in an orthonormal basis ei1≤i≤N with the matrix R = (rkl) such that vi =∑
1≤k≤N rikek. With this notation V = RR∗ and since RR∗ = (√
V )2 and we have
already shown that V is positive definite, there exists a unitary matrix U = (ukl)
defined directly by U =√
V−1
R. We write R =√
V U . Let
(6.17) wi = σ∑
1≤k≤N
uikek .
40 ILIE GRIGORESCU
We have to show that the elements of the diagonal of (R − σU)(R − σU)∗ are
uniformly bounded by a constant of order O(N−1). Since (R−σU) can be written
as σ(σ−1√
V − I)U we have
|((σ−1√
V − I)UU∗(σ−1√
V − I)ei, ei)| = ‖(σ−1√
V − I)ei‖2 ≤ ‖(σ−2V − I)ei‖2
by contraction. To see this, we denote yi = (σ−1√
V − I)ei for all 1 ≤ i ≤ N and
A = σ−1√
V + I. The positive definiteness of V implies that A−1 is a contraction.
If (δij) denotes the unit matrix, the bound for the diagonal term of R− σU is
‖(V − σ2I)ei‖2 =∑
1≤k≤N
(Vik − σ2δik)2 .
We recall that the diagonal term is of order O(N−1/2) and all the non-diagonal
terms are of order O(N−1), uniformly over the set of indices 1 ≤ i, j ≤ N . This
concludes the proof. ¤
Acknowledgements. The author would like to thank Professor S.R.S. Varadhan
for suggesting the original (harder) problem and also for simplifying the proof of
Lemma 1.
References
[1] Bogachev, Vladimir I. (1998) Gaussian measures. Mathematical Surveys and Monographs
62, American Mathematical Society, Providence, RI.
[2] Chang, C.C.; Yau, H-T. (1992) Fluctuations of one-dimensional Ginzburg-Landau models
in nonequilibrium. Commun. Math. Phys. 145, 209-234.
[3] Dunford, N. and Schwartz, J. (1988) Linear Operators, Part I, General Theory. John-Wiley
and Sons.
[4] Gielis, G., Koukkous, A., Landim, C. (1998) Equilibrium fluctuations for zero range pro-
cesses in random environment. Stochastic Process. Appl. 77, no. 2, 187–205.
[5] Grigorescu, I. (1999) Self-diffusion for Brownian motions with local interaction. Ann. Probab.
27, no. 3, 1208–1267.
[6] Grigorescu, I. (2000) Large scale behavior of a system of interacting diffusions. Hydrody-
namic limits and related topics (Toronto, ON, 1998), 83–93, Fields Inst. Commun., 27,
Amer. Math. Soc., Providence, RI.
[7] Holley, R.; Stroock, D. W. (1979) Central limit phenomena of various interacting systems.
Ann. of Math. (2) 110, no. 2, 333–393.
[8] Ito, K. (1983) Distribution-valued processes arising from independent Brownian motions.
Math. Z. 182, no. 1, 17–33.
[9] Ikeda, N. and Watanabe, S. (1989) Stochastic Differential Equations and Diffusion Processes.
Second Edition. North-Holland / Kodansha.
AN INFINITE DIMENSIONAL CENTRAL LIMIT THEOREM 41
[10] John, F. (1991) Partial differential equations. Fourth Edition. Applied Mathematical Sci-
ences, 1. Springer-Verlag, New York.
[11] Kipnis, C.; Landim, C. (1999) Scaling Limits of Interacting Particle Systems. Springer-
Verlag, New York.
[12] Quastel, J. (1998) Diffusion of color in the simple exclusion process. Commun. Pure
Appl.Math. 45, 321-379.
[13] Quastel, J.; Rezakhanlou, F.; Varadhan, S. R. S. (1999) Large deviations for the symmetric
simple exclusion process in dimensions d ≥ 3. Probab. Theory Related Fields 113, no. 1,
1–84.
[14] Ravishankar, K. (1992) Fluctuations from the hydrodynamical limit for the symmetric sim-
ple exclusion in Zd. Stochastic Process. Appl. 42, no. 1, 31–37.
[15] Shiryaev, A. N. (1984) Probability. Translated from the Russian by R. P. Boas. Graduate
Texts in Mathematics 95. Springer-Verlag, New York-Berlin.
[16] Sznitman, A.S. (1985) A fluctuation result for nonlinear diffusions. Infinite-dimensional
analysis and stochastic processes (Bielefeld, 1983), Res. Notes in Math. 124, Pitman,
Boston, Mass.-London. 145–160.
Department of Mathematics, University of Miami, 1365 Memorial Drive, Ungar Build-
ing, Room 525, Coral Gables, FL 33146
E-mail address: igrigore@math.miami.edu